What happened
The Hugging Face model card describes Alpie Seed 9B as a 9-billion-parameter multimodal model developed by 169Pi AI. It is based on Qwen3.5-9B, whose architecture the card says was retained rather than structurally redesigned. The model accepts text, images, and video, and has a stated native context length of 262,144 tokens. It supports thinking and non-thinking modes. The card lists an Apache 2.0 license and identifies the release as post-training on the Qwen3.5-9B language model.
The stated post-training process has two stages. First, supervised fine-tuning uses a DoRA adapter to improve instruction following, response quality, and task performance. Second, Group Relative Policy Optimization trains a DoRA adapter while the SFT adapter remains a frozen reference policy. For that GRPO stage, the card says the base was run in 4-bit NF4 with double quantization and BF16 computation; the trained adapter was then merged into the base model at its original precision for release. The listed architecture includes 32 layers and a 27-layer vision transformer.
The card provides examples for running the model with Transformers, vLLM, SGLang, and Docker Model Runner, and links to Ollama, Kaggle, and a separate GGUF listing. Its quickstart says support requires Transformers 5 or later and gives a Transformers 5.2-or-later installation command. The vLLM example sets the maximum model length to 262,144; an alternative command enables automatic tool-call parsing. The card explicitly notes that the application must execute requested tools and return results. Ollama tags listed include latest at 5.6 GB, q2 at 3.8 GB, q3 at 4.6 GB, and vision at 6.6 GB; the card lists a 256K context for each and specifies text-and-image input for the vision tag. It also describes a demo of a quantized version running offline on a mobile device using PocketPal.
The model card reports benchmark evaluations with extended thinking enabled, a maximum generation length of 65,536 tokens, and temperature 0.0. It gives Alpie Seed FP16 results including 87.30 on MMLU, 84.32 on GPQA Diamond, 83.33 on AIME 2026, 72.99 on LiveCodeBench v6, and 61.20 on SWE-Bench Verified. For vision-language tasks it lists 68.70 on MMMU-Pro, 80.39 on RealWorldQA, 79.88 on HallusionBench, and 90.30 on OCRBench. The card says SWE-Bench Verified uses Mini-SWE-agent with retries disabled. It also warns that comparison scores come from other published model cards and reports, were not rerun by 169Pi, and may differ in harnesses, prompts, benchmark versions, and inference settings.
What the archive adds
New Runtime's earlier coverage includes Perplexity's open-weight vision-language embedding models (2026-11-03) and Prism ML's 2B vision-language model (2026-11-07). Alpie Seed is a new model release, not a new fact about either product: it adds another open-weight multimodal model, with a different stated purpose and deployment profile.
Trend signal
This does not yet fit a pattern New Runtime tracks. A single model release with local deployment formats is not enough to establish a broader shift in how agent systems, protocols, or production workflows are changing.
Why now
The model card identifies Alpie Seed as a current release and offers downloadable weights and deployment instructions across several local and serving tools. That makes the release immediately testable by teams assessing multimodal models; the card does not give a separate explanation for the timing.
Why it matters
For teams handling long documents, images, or video, a model that can be downloaded and run through familiar serving stacks is a practical candidate for evaluation outside a hosted API. The 262K context claim and quantized options may be relevant where deployment constraints include local hardware or offline operation. But quantization changes the model being evaluated: the published scores apply to FP16, not automatically to the smaller GGUF variants.
The benchmark table is useful as a starting point, not a procurement verdict. Because the comparisons were assembled from other publishers rather than rerun under one harness, teams should compare candidates using the same prompts, task data, inference settings, and hardware. Tool-call parsing also does not itself execute tools; builders still need to implement permission checks, execution, and result handling in their application. Treat the release as an option to test against a defined workload, not as proof that it will outperform alternatives in production.
Limits and uncertainty
The benchmark results are reported by 169Pi, and the card itself cautions that cross-model comparisons are indicative rather than strictly like-for-like. It does not establish performance on a particular team's workload or hardware, and FP16 scores do not establish quality for the quantized builds. The mobile demo is described as an offline run, but the card's visible description does not quantify speed, memory use, battery draw, or task accuracy on that device. Deployment compatibility and practical quality should therefore be verified in the intended setup.
Skill to build now
Create a small evaluation set from the images, scanned documents, and long inputs your product actually handles. Run Alpie Seed FP16 and any quantized version you might deploy with identical prompts, record accuracy and latency, and note hardware and memory use. Keep tool execution outside the model: test parsed calls against an allowlist and require your application to authorize each action.
Reader takeaway
Alpie Seed 9B is an openly licensed, locally deployable multimodal model with a long stated context and multiple quantized formats. Its scores make it worth testing, but not safe to rank from the card alone: validate the exact model variant on your own workload and hardware.